Satellite signal modulation mode identification method and device and storage medium

Abstract

The invention discloses a satellite signal modulation mode identification method and device and a storage medium, and the method comprises the steps: obtaining a satellite signal, carrying out the processing of the satellite signal, obtaining a time-frequency diagram data set, and dividing the time-frequency diagram data set into a training data set, a verification data set, and a test data set; the training data set is adopted to train a neural network model, the verification data set is adopted to optimize the neural network model, a trained neural network model is obtained, and the neural network model comprises an improved mask auto-encoder and a classification head; and inputting the test data set into the trained neural network model to obtain a modulation mode of the satellite signal to be identified. The invention provides a modulation mode for recognizing satellite signals based on an improved mask auto-encoder and a neural network of a classification head, the problem of scarcity of satellite communication scene labeling samples is solved, and the recognition precision is high.

Description

Technical Field

[0001] This invention relates to the field of space spectrum cognition technology, and in particular to a method, apparatus and storage medium for identifying satellite signal modulation patterns. Background Technology

[0002] With the development of satellite technology, satellite constellation systems are gradually entering the stage of large-scale deployment. Among these technologies, automatic identification of satellite signal modulation modes is a core technology for satellite signal detection and demodulation, requiring the receiver to quickly and accurately identify the modulation mode of satellite signals even without prior knowledge. This technology is also a key component of low-Earth orbit satellite spectrum monitoring systems, and the accuracy and efficiency of its identification and classification directly affect the optimal allocation of spectrum resources.

[0003] Automatic identification of satellite signal modulation schemes has developed into two main technical approaches: traditional identification methods and deep learning-driven identification algorithms. Traditional identification methods mainly include maximum likelihood hypothesis testing based on decision theory and pattern recognition based on feature extraction. However, these methods only perform identification on a single parameter, and the extracted signal features in low signal-to-noise ratio environments lack sufficient discriminative power, resulting in a low overall identification rate. Deep learning technology, on the other hand, possesses a powerful fitting capability for handling complex nonlinear problems. Introducing it into the field of signal identification can significantly improve the accuracy of modulation scheme identification in satellite communication systems.

[0004] Current deep learning-based recognition methods suffer from several problems: scarce labeled samples and limited receptive fields in convolutional neural networks. Therefore, it is necessary to develop a novel deep learning-based recognition method to accurately identify the modulation patterns of satellite signals. Summary of the Invention

[0005] This invention aims to at least partially solve one of the technical problems in related technologies. To this end, one objective of this invention is to provide a method, apparatus, and storage medium for identifying satellite signal modulation patterns. It proposes a neural network based on an improved mask autoencoder and a classification head to address the problem of scarce labeled samples in satellite communication scenarios. Furthermore, the features extracted through the mask autoencoder are comprehensive, resulting in high recognition accuracy.

[0006] The technical solution adopted in this invention is: In a first aspect, the present invention provides a method for identifying the modulation scheme of a satellite signal, comprising: acquiring a satellite signal, processing the satellite signal to obtain a time-frequency map dataset, and dividing the time-frequency map dataset into a training dataset, a validation dataset, and a test dataset; training a neural network model using the training dataset, optimizing the neural network model using the validation dataset to obtain a trained neural network model, the neural network model including an improved mask autoencoder and a classification head; and inputting the test dataset into the trained neural network model to obtain the modulation scheme of the satellite signal to be identified.

[0007] The process of acquiring satellite signals and processing them to obtain a time-frequency map dataset, and then dividing the time-frequency map dataset into a training dataset, a validation dataset, and a test dataset, includes: acquiring and processing raw satellite signals to obtain standardized satellite signals; denoising the standardized signals using a multi-wavelet basis thresholding denoising method; generating a time-frequency map by time-frequency transformation and optimization of the denoised signals; and dividing the time-frequency map into the training dataset, the validation dataset, and the test dataset according to a preset ratio.

[0008] The improved mask autoencoder includes an encoder and a decoder. The encoder employs a lightweight Transformer encoder, which includes a Linear Embedding module, a first Reshape module, a first STL module, and a second Reshape module. The decoder employs a lightweight Vision Transformer structure, which includes a dual-signal-stream fusion module, a third Reshape module, a second STL module, and a fourth Reshape module. The dual-signal-stream fusion module introduces an attention mechanism to integrate the two streams of data in the latent space. The masking strategy of the mask autoencoder is designed to prioritize encoding blocks with higher frequency components. The loss function uses the mean squared error function, and the Adam optimizer is used during training. The weight decay is 0.0001, the initial learning rate is 0.001, the batch size is adjustable, and the learning rate is scheduled using a cosine annealing strategy to avoid overfitting and accelerate convergence. Based on the pre-trained encoder and decoder, the model is fine-tuned using a small number of labeled samples.

[0009] The method of denoising the standardized signal using a multi-wavelet basis thresholding method includes: performing multi-level decomposition of the standardized signal using the Mallat algorithm; performing Heursure adaptive thresholding on the decomposed signal to obtain the denoised signal; and performing median filtering smoothing on the denoised signal to eliminate residual noise after wavelet transform denoising.

[0010] The process of acquiring and processing raw satellite signals to obtain standardized satellite signals includes: receiving satellite radio frequency signals, down-converting and sampling them to obtain discrete signals; converting the discrete signals into a two-dimensional IQ signal matrix; and performing amplitude normalization processing on the two-dimensional IQ signals to generate the standardized satellite signals.

[0011] Secondly, the present invention provides a satellite signal modulation mode identification device, the identification device comprising: a satellite signal acquisition and processing module, used to acquire satellite signals, process the satellite signals to obtain a time-frequency map dataset, and divide the time-frequency map dataset into a training dataset, a validation dataset, and a test dataset; a neural network model training module, used to train a neural network model using the training dataset, and optimize the neural network model using the validation dataset to obtain a trained neural network model, the neural network model including an improved mask autoencoder and a classification head; and a satellite signal modulation mode identification module, used to input the test dataset into the trained neural network model to obtain the modulation mode of the satellite signal to be identified.

[0012] The satellite signal acquisition and processing module includes: a satellite signal acquisition and processing unit for acquiring and processing raw satellite signals to obtain standardized satellite signals; a satellite signal denoising unit for denoising the standardized signal using a multi-wavelet basis threshold denoising method; a satellite signal conversion unit for generating a time-frequency map from the denoised signal through time-frequency transformation and optimization; and a time-frequency map partitioning unit for partitioning the time-frequency map into the training dataset, the validation dataset, and the test dataset according to a preset ratio.

[0013] The satellite signal denoising unit includes: a satellite signal decomposition subunit, used to perform multi-level decomposition of the standardized signal using the Mallat algorithm; a satellite signal denoising subunit, used to perform Heursure adaptive threshold calculation on the decomposed signal to obtain the denoised signal; and a residual signal denoising subunit, used to perform median filtering smoothing on the denoised signal to eliminate residual noise after wavelet transform denoising.

[0014] The satellite signal acquisition and processing unit includes: a satellite signal acquisition subunit, used to receive satellite radio frequency signals and obtain discrete signals through down-conversion and sampling; a satellite signal conversion subunit, used to convert the discrete signals into a two-dimensional IQ signal matrix; and a satellite signal normalization processing subunit, used to perform amplitude normalization processing on the two-dimensional IQ signals to generate the standardized satellite signal.

[0015] Thirdly, the present invention provides a computer-readable storage medium storing computer-executable instructions for causing a computer to perform the method described above.

[0016] The beneficial effects of this invention are: This invention proposes a neural network based on an improved masked autoencoder and a classification head to identify the modulation scheme of satellite signals. This method enables deep learning-based modulation scheme identification even when satellite signal samples are scarce, thereby reducing signal annotation costs and addressing the problem of limited labeled samples in satellite communication scenarios. Furthermore, this invention uses time-frequency maps as input to the neural network model, and the unsupervised pre-training of the masked autoencoder enhances the global feature extraction capability of the time-frequency map, thus improving recognition accuracy.

[0017] Furthermore, this invention addresses the problem of mask autoencoder reconstruction failure caused by noise distortion of the original satellite signal during the satellite signal processing stage by using continuous wavelet transform and multi-wavelet basis denoising.

[0018] Furthermore, the mask autoencoder of this invention employs a dual-signal-stream fusion module and introduces an attention mechanism fusion strategy, effectively enhancing the robustness and recognition accuracy of the model under low signal-to-noise ratio conditions. Simultaneously, the mask autoencoder's masking strategy designs a block-priority mask with multiple frequency components, avoiding continuous masking of the same frequency component, preserving the key frequency features of the satellite modulation signal, making the reconstruction task more aligned with modulation recognition requirements, and enhancing the discriminative power of the pre-trained features.

[0019] In addition, this invention optimizes the time-frequency diagram and mask autoencoder to reduce the number of parameters in the overall network framework and adapt to the resource constraints of the satellite platform.

[0020] Furthermore, this invention normalizes the amplitude of the acquired satellite signals, mapping the amplitude of all sampling points to the interval [−1,1], thereby eliminating the impact of signal strength differences on model training. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating an embodiment of the satellite signal modulation mode identification method of the present invention; Figure 2 yes Figure 1 A flowchart illustrating an embodiment of step S11; Figure 3 yes Figure 2 A flowchart illustrating an embodiment of step S111; Figure 4 yes Figure 2 A flowchart illustrating an embodiment of step S112; Figure 5 yes Figure 2 A schematic diagram of the structure of an embodiment of step S112; Figure 6 yes Figure 2A schematic diagram of the structure of an embodiment of step S113; Figure 7 yes Figure 1 A schematic diagram of the structure of an embodiment of the neural network model in step S12; Figure 8 This is a schematic diagram of an embodiment of the satellite signal modulation mode identification device of the present invention; Figure 9 yes Figure 8 A schematic diagram of the structure of an embodiment of the satellite signal acquisition and processing module 11; Figure 10 yes Figure 9 A schematic diagram of the structure of an embodiment of the satellite signal acquisition and processing unit 111; Figure 11 yes Figure 10 A schematic diagram of an embodiment of the satellite signal denoising unit 112. Detailed Implementation

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0023] Example 1

[0024] Please see Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the satellite signal modulation scheme identification method of the present invention. Figure 1 As shown, the identification method includes the following steps: S11: Acquire satellite signals, process the satellite signals to obtain a time-frequency map dataset, and divide the time-frequency map dataset into a training dataset, a validation dataset, and a test dataset; The modulation methods for signals in satellite communication systems include the following types: 2ASK, 4ASK, 2PSK, 4PSK, 8PSK, 2FSK, 4FSK, 16QAM, and 64QAM. ASK is amplitude shift keying, PSK is phase shift keying, FSK is frequency shift keying, and QAM is quadrature amplitude modulation.

[0025] Please see Figure 2 ,like Figure 2 As shown, step S11 includes the following sub-steps: S111: Acquire and process raw satellite signals to obtain standardized satellite signals; Please see Figure 3 ,like Figure 3 As shown, step S111 includes the following sub-steps: S1111: Receives satellite radio frequency signals, and obtains discrete signals through down-conversion and sampling; The receiver receives satellite radio frequency signals, converts them to intermediate frequency signals via down-conversion, and then samples them using an analog-to-digital converter (ADC) to obtain discrete signals. The sampling rate is adapted to the carrier frequency to ensure that the Nyquist sampling theorem is satisfied, and the sampling frequency... It must be greater than the highest frequency of the signal. Twice as of, that is: .

[0026] The raw satellite signals acquired in this step can be represented by the following mathematical model: ,in, The original signal, For fading channel gain, For modulation signal, It is zero-mean Gaussian white noise.

[0027] S1112: Convert the discrete signal into a two-dimensional IQ signal matrix; The two-dimensional IQ signal matrix is ​​represented as follows: The two channels correspond to the in-phase component I and the quadrature component Q, respectively, and 1024 is the number of sampling points for each channel.

[0028] This step can be represented by the following mathematical model: ,in, The original signal collected at each sampling point, It is the sampling interval. The Doppler frequency shift is randomly distributed between -500Hz and 500Hz.

[0029] S1113: Perform amplitude normalization processing on each IQ signal of the two-dimensional IQ signal matrix to generate the standardized satellite signal.

[0030] The amplitude of each IQ signal in the two-dimensional IQ signal matrix is ​​normalized, mapping the amplitude of all sampling points to the interval [−1, 1], thus eliminating the influence of signal intensity differences on model training. The normalization formula is: ,in, These are the original sampled values. , These represent the maximum and minimum amplitudes of the signal, respectively.

[0031] S112: Denoise the standardized signal using a multi-wavelet basis threshold denoising method; Please refer to the following: Figure 4 and Figure 5 ,like Figure 4 As shown, step S112 includes the following steps: S1121: The normalized signal is decomposed into multiple levels using the Mallat algorithm; The specific decomposition process is as follows: First, the original signal undergoes a first-level wavelet decomposition to obtain an approximate signal a1 and a detail signal b1 containing high-frequency noise. Next, a second-level wavelet decomposition is performed on signal a1 using two different wavelet bases to obtain a low-frequency approximate signal a2 and a high-frequency signal b2. Finally, a2 is decomposed to obtain a low-frequency approximate signal a3 and a high-frequency detail signal b3. The Mallat algorithm is then used to obtain the final-level approximate signal a3 and the detail signals b1, b2, and b3.

[0032] The three-level decomposition described above all use the Mallat algorithm. In other embodiments, four or more levels of decomposition may also be used.

[0033] S1122: Perform Heursure adaptive threshold calculation on the decomposed signal to obtain the denoised signal; The Heursure threshold selection rule is used to determine the denoising thresholds for signals a3, b1, b2, and b3. The Sqtwolog threshold and the Rigrsure threshold are then combined to adaptively determine the denoising thresholds for detail signals at each level. The formula is as follows: , in, , To minimize the risk threshold, , In the formula For the first Hierarchical signal length.

[0034] By using thresholds to shrink and reconstruct signals b1, b2, and b3, noise components are suppressed. The specific expression is as follows: ,in It is a noise-reduced detail signal wavelet coefficient set, As the sign function, through the above transformation, the final reconstructed signal is obtained as follows: .

[0035] S1123: Perform median filtering to smooth the denoised signal and eliminate residual noise after wavelet transform denoising.

[0036] Since the approximate signal a3 after wavelet denoising in step S1122 still exhibits non-uniformity, a median filter of order 10 is applied to smooth the signal, thereby eliminating residual impulse noise after wavelet transform denoising. The specific formula is as follows: ; in, This indicates removing the median filter from the samples within the window, covering 10 points. These are the denoised discrete signals, respectively.

[0037] S113: Generate a time-frequency diagram by performing time-frequency transformation and optimization on the denoised signal; Please see Figure 6 In this step, the short-time Fourier transform (STFT) can be used to convert the graph into a two-dimensional time-frequency diagram, specifically represented as follows: Among them, time-frequency images H and W are used to represent the time axis and frequency axis resolutions of the generated time-frequency plot, respectively. More specifically, the original IQ signal in complex form (containing in-phase component I and quadrature component Q) is converted into a time-frequency plot using a short-time Fourier transform (STFT). After STFT processing, the original IQ signal is converted into a two-dimensional time-frequency plot, where the time axis length H is determined by the window length and overlap rate, and the frequency axis length W is half the number of Fourier transform points.

[0038] After being converted into a time-frequency graph, a center frequency determination method is used to select the effective frequency components around the center frequency component, while the time component can be downsampled to generate a compact time-frequency graph, thereby reducing the size of the input image and the amount of training required for the model.

[0039] The final output time-frequency graph has a size of 256×256×1, which serves as the input data for the subsequent neural network model.

[0040] S114: Based on the preset ratio, divide the time-frequency graph into the training dataset, the validation dataset, and the test dataset.

[0041] Preferably, the preset ratio is 8:1:1. Of course, other ratios can also be set.

[0042] S12: Use the training dataset to train the neural network model, and use the validation dataset to optimize the neural network model to obtain a trained neural network model; like Figure 7 As shown, this neural network model includes an improved masked autoencoder and a classification head. The classification head employs the softmax function.

[0043] The configuration used in step S12 for training the neural network model can be found in Table 1 below: Table 1 Training Part Parameter Configuration Table

[0044] This improved mask autoencoder is based on the MAE framework, which includes an encoder and a decoder.

[0045] The encoder employs a lightweight Transformer encoder, comprising a Linear Embedding module, a first Reshape module, a first STL module, and a second Reshape module. The Linear Embedding module linearly maps the time-frequency image data and adds positional encoding information to obtain shallow information of the time-frequency image. The first Reshape module performs data format transformation. The first STL module acquires deep information of the time-frequency image. The second Reshape module extracts global features from the unmasked patches. The encoder's structural parameters are adapted to satellite time-frequency images. The neural network has six stacked layers to balance feature extraction capability and computational cost; it uses eight attention heads, each with eight dimensions, for a total of 64 dimensions; the GELU activation function mitigates gradient vanishing; each unmasked patch is linearly projected onto 64-dimensional features, combined with learnable positional encoding, to obtain the input features. .

[0046] The decoder employs a lightweight Vision Transformer structure to reconstruct the time-frequency features of the mask patch and supervise the encoder's learning. The decoder has four layers, forming an asymmetric structure with the encoder; the number of layers is less than six to avoid overfitting. It also has four attention heads, each with eight dimensions, for a total of 32 dimensions. The encoder output is then processed... The input to the decoder is obtained by concatenating the mask positions with the input. Finally, the reconstructed features of each patch output by the decoder are mapped back to the patch dimension of the time-frequency map through linear projection to obtain the final reconstructed time-frequency map. The decoder includes a dual-signal-stream fusion module, a third reshape module, a second STL module, and a fourth reshape module. The dual-signal-stream fusion module is used to integrate the two data streams in the latent space using an attention mechanism fusion strategy. The third reshape module handles data transformation. The second STL module obtains deeper data image information. The fourth reshape module restores the shape of the time-frequency map after reshape transformation and performs linear probability classification on the restored time-frequency map to determine the type of modulated time-frequency map.

[0047] The Dual-stream Signal Fusion Module (DSSFM) is a processing module built upon wavelet transform denoising and time-frequency transform. Its core design philosophy lies in achieving multi-level feature extraction and deep fusion of complex signals through a parallelized dual-channel architecture. The attention matrix of the STL structure within this module is represented as follows: ; in, Let represent the query, key, and value, respectively; let represent the learnable relative position encoding; and d represent the dimension of the multi-head attention. Next, layer normalization (LN) is added between the MSA and the multi-layer perceptron (MLP) using the GULE activation function and residual connections for further feature enhancement.

[0048] This module first feeds the input signal into two independent processing streams: one is a wavelet decomposition stream, which decomposes the original signal into sub-band components of different frequency bands through multi-scale wavelet transform, accurately capturing the local features and transient changes in the time and frequency domains of the signal; the other is an adaptive feature enhancement stream, which combines a deep learning network to perform nonlinear mapping and dimensional expansion on the key coefficients after wavelet decomposition, enhancing the representation ability of high-frequency detail features and low-frequency trend features. In the feature fusion stage, DSSFM adopts a dynamic weight allocation mechanism, adaptively adjusting the fusion weights according to the signal-to-noise ratio and information entropy of each sub-band signal, focusing on important feature components through an attention mechanism, and finally achieving the collaborative reconstruction of multi-dimensional features through inverse wavelet transform, effectively improving the anti-interference capability and feature recognition accuracy of the signal processing system.

[0049] After determining the neural network framework, the loss function and optimization strategy for training are designed. Since training is performed on a time-frequency map, the objective is to "reconstruct the time-frequency map of the masked region," and the mean squared error (MSE) loss function is used, calculating only the reconstruction error of the masked region. ; in, For the first The reconstruction result of each patch, P i This represents the i-th patch block. This is the i-th masked patch block. During training, the Adam optimizer is used with a weight decay of 0.0001 and an initial learning rate of 0.001. The batch size is adjustable, and the training epochs are approximately two hundred. The learning rate is scheduled using a cosine annealing strategy to avoid overfitting and accelerate convergence.

[0050] Based on the pre-trained encoder and decoder, the model is fine-tuned using a small number of labeled samples. During feature extraction, the parameters of the first four layers of the encoder are frozen to retain the general time-frequency features of unsupervised learning. The parameters of the last two layers are fine-tuned, and the encoder outputs global features. The mean of the global features is taken to compress the dimensionality and retain global correlations. A classification network is constructed using a two-layer fully connected network to output the probability distribution of the modulation scheme. The classification probability is obtained using the Softmax function.

[0051] Preferably, the masking strategy of the masking autoencoder is designed to prioritize masking blocks with more frequency components. Before masking, the time-frequency image contains many components with little frequency information. Therefore, the information content of the image patch block is determined based on the chromaticity information of the time-frequency image. When masking the time-frequency image, blocks with more frequency components are masked first to avoid continuously masking the same frequency component, thus preserving the key frequency features of the satellite modulation signal. This makes the reconstruction task more in line with the modulation recognition requirements and the pre-trained features more discriminative.

[0052] S13: Input the test dataset into the trained neural network model to obtain the modulation mode of the satellite signal to be identified; This step can automatically output the modulation mode of the satellite signal to be identified, whether it is 2ASK, 4ASK, 2PSK, 4PSK, 8PSK, 2FSK, 4FSK, 16QAM, or 64QAM.

[0053] The test time-frequency graph is input into the trained neural network model to obtain the probability of each of the above modulation methods. The modulation method corresponding to the maximum probability is taken as the modulation method of the satellite signal to be identified.

[0054] This embodiment addresses the problems of scarce samples, weak generalization ability, limited receptive field of convolutional neural networks, and large training computation in deep learning-based low-Earth orbit satellite signal modulation recognition methods. It proposes a targeted optimization method: employing a neural network based on the fusion of an improved masked autoencoder and a classification head, using a wavelet-enhanced time-frequency graph as model input, and integrating the noise suppression characteristics of wavelet transform, the unsupervised pre-training advantages of the MAE framework, and the global perception capability of the Transformer to achieve high-precision modulation recognition in complex satellite communication environments. Simultaneously, an attention mechanism fusion strategy is introduced to effectively enhance the model's robustness and recognition accuracy under low signal-to-noise ratio conditions, further improving the model's adaptability in different satellite communication scenarios.

[0055] Example 2 Please see Figure 8 , Figure 8 This is a schematic diagram of an embodiment of the satellite signal modulation mode identification device of the present invention. Figure 8 As shown, the identification device includes a satellite signal acquisition and processing module 11, a neural network model training module 12, and a satellite signal modulation mode identification module 13.

[0056] The satellite signal acquisition and processing module 11 is used to acquire satellite signals, process the satellite signals to obtain a time-frequency map dataset, and divide the time-frequency map dataset into a training dataset, a validation dataset, and a test dataset. The neural network model training module 12 is used to train the neural network model using the training dataset and to optimize the neural network model using the validation dataset, thereby obtaining a trained neural network model, which includes an improved mask autoencoder and a classification head.

[0057] The satellite signal modulation mode identification module 13 is used to input the test dataset into the trained neural network model to obtain the modulation mode of the satellite signal to be identified.

[0058] Please see Figure 9 The satellite signal acquisition and processing module 11 includes a satellite signal acquisition and processing unit 111, a satellite signal denoising unit 112, a satellite signal conversion unit 113, and a time-frequency map partitioning unit 114. The satellite signal acquisition and processing unit 111 acquires and processes the raw satellite signal to obtain a standardized satellite signal. The satellite signal denoising unit 112 denoises the standardized signal using a multi-wavelet basis thresholding method. The satellite signal conversion unit 113 generates a time-frequency map from the denoised signal through time-frequency transformation and optimization. The time-frequency map partitioning unit 114 divides the time-frequency map into a training dataset, a validation dataset, and a test dataset according to a preset ratio.

[0059] Please see Figure 10 The satellite signal acquisition and processing unit 111 includes a satellite signal acquisition subunit 1111, a satellite signal conversion subunit 1112, and a satellite signal standardization processing subunit 1113. The satellite signal acquisition subunit 1111 receives satellite radio frequency signals and obtains discrete signals through down-conversion and sampling. The satellite signal conversion subunit 1112 converts the discrete signal into a two-dimensional IQ signal matrix. The satellite signal standardization processing subunit 1113 performs amplitude normalization processing on the two-dimensional IQ signal to generate the standardized satellite signal.

[0060] Please see Figure 11 The satellite signal denoising unit 112 includes a satellite signal decomposition subunit 1121, a satellite signal denoising subunit 1122, and a residual signal denoising subunit 1123. The satellite signal decomposition subunit 1121 performs multi-level decomposition of the normalized signal using the Mallat algorithm. The satellite signal denoising subunit 1122 performs Heursure adaptive thresholding on the decomposed signal to obtain the denoised signal. The residual signal denoising subunit 1123 performs median filtering smoothing on the denoised signal to eliminate residual noise after wavelet transform denoising.

[0061] Specifically, the working methods of each module in this embodiment have been described in detail in Embodiment 1, and will not be repeated here.

[0062] Example 3 The present invention also provides a computer-readable storage medium storing computer-executable instructions for causing a computer to perform the method described in Embodiment 1.

[0063] In summary, this invention proposes a neural network based on the fusion of wavelet enhancement and improved masked autoencoder for satellite signal modulation pattern recognition. In the data preprocessing stage, continuous wavelet transform and multi-wavelet basis denoising are used to suppress noise interference and address feature distortion issues under low signal-to-noise ratio. Then, high-quality time-frequency maps are generated using the denoised signal data. In terms of data model construction, unsupervised pre-training of MAE is introduced, utilizing unlabeled satellite signals to pre-train a feature extractor, reducing the need for labeled samples in the fine-tuning stage, while simultaneously enhancing global feature capture capabilities. The MAE-pretrained Transformer encoder is combined with a classification head to fully exploit global correlations in the time-frequency map, improving the ability to distinguish between multiple types of modulation signals. In engineering, end-to-end integration is achieved, constructing a fully optimized modulation recognition framework from satellite signal acquisition, wavelet preprocessing, MAE pre-training to fine-tuning and classification, thus improving engineering deployment efficiency.

[0064] By employing this invention, the level of signal modulation mode identification in satellite communication systems can be improved, and the system's adaptability to complex environments can be enhanced, thus contributing to the monitoring and management of satellite spectrum resources.

[0065] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method of identifying a modulation scheme of a satellite signal, characterized by, The method comprises the following steps: acquiring satellite signals, processing the satellite signals to obtain a time-frequency graph data set, and dividing the time-frequency graph data set into a training data set, a verification data set, and a test data set; training a neural network model using the training data set, optimizing the neural network model using the verification data set, and obtaining a trained neural network model, wherein the neural network model comprises an improved mask autoencoder and a classification head; inputting the test data set into the trained neural network model to obtain a modulation mode of a satellite signal to be identified.

2. The identification method according to claim 1, characterized in that, The method of acquiring satellite signals and processing the satellite signals to obtain a time-frequency graph data set, and dividing the time-frequency graph data set into a training data set, a verification data set, and a test data set comprises the following steps: acquiring and processing original satellite signals to obtain standardized satellite signals; denoising the standardized signals using a multi-wavelet basis threshold denoising method; generating a time-frequency graph through time-frequency transformation and optimization on the denoised signals; dividing the time-frequency graph into the training data set, the verification data set, and the test data set according to a preset ratio.

3. The identification method according to claim 1, characterized in that, The improved mask autoencoder comprises an encoder and a decoder, the encoder adopts a lightweight Transformer encoder, the encoder comprises a Linear Embedding module, a first Reshape module, a first STL module, and a second Reshape module, the decoder adopts a lightweight vision Transformer structure, the decoder comprises a double-signal stream fusion module, a third Reshape module, a second STL module, and a fourth Reshape module, and the double-signal stream fusion module is used to introduce an attention mechanism fusion strategy to integrate two-way data in a hidden space. The mask strategy of the mask autoencoder is designed to preferentially encode blocks with more frequency components. The loss function adopts a mean square error function, the Adam optimizer is used for training, the weight decay is 0.0001, the initial learning rate is 0.001, the batch size is adjustable, the learning rate is scheduled through a cosine annealing strategy to avoid overfitting and accelerate convergence. Based on the pre-trained encoder-decoder, the model is fine-tuned using a small number of labeled samples.

4. The identification method according to claim 2, characterized in that, The method of denoising the standardized signals using a multi-wavelet basis threshold denoising method comprises the following steps: performing multi-level decomposition on the standardized signals using a Mallat algorithm; calculating a Heursure adaptive threshold for the decomposed signals to obtain denoised signals; performing median filtering and smoothing on the denoised signals to eliminate residual noise after wavelet transform denoising.

5. The identification method according to claim 2, characterized in that, The method of acquiring and processing original satellite signals to obtain standardized satellite signals comprises the following steps: receiving satellite radio frequency signals, down-converting and sampling to obtain discrete signals; converting the discrete signals into a two-dimensional IQ signal matrix; performing amplitude normalization processing on the two-dimensional IQ signals to generate the standardized satellite signals.

6. An apparatus for identifying a modulation scheme of a satellite signal, characterized by The method comprises the following steps: The satellite signal acquisition and processing module is configured to acquire satellite signals, process the satellite signals to obtain a time-frequency diagram data set, and divide the time-frequency diagram data set into a training data set, a verification data set, and a test data set. The neural network model training module is configured to train a neural network model using the training data set, optimize the neural network model using the verification data set, and obtain a trained neural network model, wherein the neural network model comprises an improved mask autoencoder and a classification head. The satellite signal modulation mode identification module is configured to input the test data set into the trained neural network model to obtain a modulation mode of a satellite signal to be identified.

7. The identification device of claim 6, wherein The satellite signal acquisition and processing module comprises: A satellite signal acquisition and processing unit configured to acquire and process original satellite signals to obtain standardized satellite signals. A satellite signal denoising unit configured to denoise the standardized signals using a multi-wavelet basis threshold denoising method. A satellite signal conversion unit configured to generate a time-frequency diagram through time-frequency conversion and optimization of the denoised signals. A time-frequency diagram division unit configured to divide the time-frequency diagram into the training data set, the verification data set, and the test data set according to a preset ratio.

8. The identification device of claim 7, wherein, The satellite signal denoising unit comprises: A satellite signal decomposition subunit configured to perform multi-level decomposition on the standardized signals using a Mallat algorithm. A satellite signal denoising subunit configured to perform Heursure adaptive threshold calculation on the decomposed signals to obtain denoised signals. A residual signal denoising subunit configured to perform median filter smoothing on the denoised signals to eliminate residual noise after wavelet transform denoising.

9. The identification device of claim 7, wherein, The satellite signal acquisition and processing unit comprises: A satellite signal acquisition subunit configured to receive satellite radio frequency signals, down-convert and sample the signals to obtain discrete signals. A satellite signal conversion subunit configured to convert the discrete signals into a two-dimensional IQ signal matrix. A satellite signal standardization processing subunit configured to perform amplitude normalization processing on the two-dimensional IQ signals to generate the standardized satellite signals.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer executable instructions for causing a computer to execute the method of any one of claims 1 to 5.

Images